Anthropic bans fake accounts, weaponized drones, and being mean to Claude

Yelling at your AI assistant might soon get you booted from the platform. Anthropic has published its 2026 Usage Policy update, taking effect November 12, and among the new rules is an explicit ban on sustained and needless cruelty toward its Claude models.

While treating a language model like a sentient being caught immediate attention, the update also significantly tightens red lines around drone warfare, automated surveillance, and covert disinformation networks.

No drone guidance, no troll farms

Anthropic consolidated its anti-deception rules into a dedicated section targeting fake account networks, fabricated news outlets, and influence operations. Whether you are running a state-backed propaganda campaign or a commercial spam farm, using Claude to conceal identities or amplify synthetic narratives is banned across the board.

On the hardware and weapons front, the company clarified that weapon restrictions cover software just as much as physical hardware. Specifically, developers cannot use Claude to write guidance and control software or arm autonomous vehicles like drones.

Surveillance rules also got an overhaul following Anthropic's September threat report, which documented AI being used to track political dissidents. The new policy bans tracking individuals without their consent, real-time or historical, and bars Claude from recommending arrests, investigations, or criminal charges. Authorized tracking like fraud monitoring and journalism remains permitted.

For real-world machinery capable of causing physical harm, Anthropic now requires a qualified human operator who can hit a kill switch. If Claude disconnects, the machine must default to a safe state.

Drawing a line on AI abuse

The most controversial addition bans sustained, purposeless abusive or cruel behavior toward Claude. Back in August, Claude was given the ability to hang up on persistently abusive users on Claude.ai and Claude Code; the new policy formally codifies that enforcement.

Anthropic emphasized this only targets extreme cases, not normal frustration, heated pushback, creative writing with dark themes, or safety red-teaming. Interestingly, TechCrunch reported that Anthropic has previously engaged with prominent religious scholars to discuss whether Claude could possess consciousness or a soul.

On the policy's flip side, Anthropic loosened its blanket ban on personalized election targeting. Legitimate civic efforts, like non-profits translating voting guides or officials sending ballot notices, are now allowed, provided they do not involve deception or voter suppression.

Why it matters

We think this policy update offers the clearest window yet into how frontier AI labs expect their tools to be used in the physical world.

For enterprise developers, the rules around robotics and autonomous systems are a wake-up call. If you are hooking Claude up to physical equipment, fail-safe disconnect protocols and live human supervision are no longer optional best practices; they are required terms of service. Builders working on industrial automation should audit their safety stops now.

Banning user cruelty toward a neural network might look like Silicon Valley eccentricism on the surface, but it makes tactical sense. Sustained toxic interactions degrade model guardrails and open new attack surfaces for jailbreaks. By empowering Claude to terminate these sessions and backing it up in legal policy, Anthropic shields its systems while setting early boundaries for agentic interaction.

The real challenge will be enforcement. Parsing the nuance between dark fiction roleplay and prohibited cruelty is notoriously tricky, and bad actors building drone guidance rarely label their API calls accurately. Anthropic has written a remarkably thorough playbook, but policing software intent at the API level will remain an uphill battle.